A strategyproof deep RL mechanism reached 97.8 percent of the offline welfare bound on a resource-allocation stream, against 93.8 percent for a linear rule and 89.0 percent for first-come first-served
AspireAI community · arXiv
Online resource allocation has two coupled problems: allocating scarce capacity efficiently under uncertain demand, and making sure participants report their valuations truthfully. Existing strategyproof reinforcement learning restricts itself to a linear allocation rule, while neural mechanism design is expressive but only approximately incentive-compatible. This paper puts the two together with a monotonic neural network that learns nonlinear interactions by construction, so a higher reported value cannot reduce acceptance and a longer reported duration or larger resource requirement cannot increase it, plus a generalised critical payment defined as the lowest accepted value with all other inputs fixed.
On a four-resource stream with two partner pairs - a case a linear difference cannot represent - the neural rule reaches 97.8 percent of an offline welfare upper bound against 93.8 percent for a linear learner and 89.0 percent for first-come first-served, closing 65 and 80 percent of the remaining optimality gap. Trained by imitation of an existing rule it reaches 97.5 percent in 261 seconds, about 89 times faster than reinforcement learning.
The attack analysis is what makes this a security story as much as an allocation story. At the most exploitable checkpoint, 43 percent of the audited states of an unconstrained MLP and 29 percent of a value-MLP admit a profitable lie, and the mechanism then drives those numbers to zero: every monotone score stays at zero regret on every stream, including the most selective partner checkpoint at 18 percent acceptance. The honest limits are the scale of the evaluation, a simulated stream rather than a deployed marketplace, and a strategyproofness guarantee that depends on stated complexity and optimisation assumptions.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.